Smart marketing system, device and medium

Abstract

The invention discloses a smart marketing system and device and a medium, and belongs to the field of big data analysis. The system comprises a data support layer used for collecting multi-source heterogeneous data including equipment data, user behavior data and third-party data, and forming data assets by preprocessing the multi-source heterogeneous data; the intelligent decision-making layer is used for identifying the life cycle stage of the current user based on an intelligent model matrix and the data assets, predicting the transition stage of the life cycle of the user and generating a precision marketing strategy and a channel selection scheme of each stage; and the business application layer is used for converting the precision marketing strategy and the channel selection scheme into a specific marketing task, monitoring marketing effect data and feeding back the marketing effect data to the data support layer and the intelligent decision-making layer. According to the system, the intelligent marketing process of full-link data fusion, dynamic life cycle management and automatic marketing closed loop can be realized.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of big data analysis, and particularly relates to a smart marketing system, device and medium. BACKGROUND

[0002] With the continuous improvement of enterprise user operation refinement requirements, the traditional extensive marketing mode is difficult to meet the business growth demand. The traditional marketing method usually relies on static user data (such as user basic attributes, transaction records) and manual operation, and uniformly pushes standardized marketing content (such as holiday coupons, new product notifications) to different groups, or uses the same marketing strategy for a long time for the same group, which is difficult to respond to multi-source data and dynamic user behavior.

[0003] With the development of big data technology and machine learning, smart marketing based on big data and user life cycle has become an important direction for industry upgrading, but the existing technology is rigid in managing the user life cycle, and most schemes can only make rough stage division, and the division rules are heavily dependent on manual experience or static indicators, which cannot accurately identify based on user behavior changes, the association between marketing strategies and the current actual stage and behavior intention of the user is not high, the strategy library is updated slowly, and personalized optimization of marketing strategies cannot be achieved, at the same time, the selection of marketing channels is mostly based on experience, lacking data-driven real-time optimization, more importantly, from the generation of marketing strategies to execution, to effect evaluation and feedback, the entire data link is broken.

[0004] The related existing patents are analyzed as follows: CN119515466A discloses a multi-platform internet marketing activity effect evaluation and prediction model construction method, which can realize multi-platform data fusion and effect prediction, but focuses on post-evaluation and effect prediction, does not deeply combine the dynamic stage of the user life cycle, lacks the ability to generate strategies based on user stage characteristics, and cannot realize the closed loop of user stage identification-targeted marketing-effect feedback.

[0005] CN115563301A discloses a user life cycle prediction method, which focuses on the life cycle length prediction of financial customers for customer maintenance, does not involve marketing action design and execution based on life cycle stages, and the data sources are limited to financial related features (such as assets, liabilities), without the fusion of multi-source non-financial data (such as device behavior data), and the application scenario is single.

[0006] CN113297478A discloses an information pushing method based on user life cycle, which can generate marketing information based on user stages, but the granularity of user grouping and strategy generation is coarse, relies on static core indicators (such as repeat purchase rate, growth rate), does not dynamically adjust the strategy in combination with real-time user behavior data, and lacks a model optimization mechanism driven by marketing effect feedback.

[0007] CN118365362A discloses a user life cycle value prediction method, focuses on improving user life cycle value (LTV) prediction accuracy through missing state coding, mainly serves game advertisement delivery, does not build a full-link marketing closed loop covering data collection, stage identification, strategy execution and effect evaluation, and has insufficient adaptability to non-internet scenes, and is difficult to support cross-industry intelligent marketing needs.

[0008] In summary, the existing marketing system has technical defects such as data island, static management, broken data link, and poor adaptability, therefore, there is an urgent need for an intelligent marketing solution that can realize full-link data fusion, dynamic life cycle management, and automatic marketing closed loop. SUMMARY

[0009] To solve the above problems, the present application provides an intelligent marketing system based on big data analysis and user life cycle, which can realize the intelligent marketing process of full-link data fusion, dynamic life cycle management and automatic marketing closed loop.

[0010] The technical solution is as follows: An intelligent marketing system comprises: A data support layer is used to collect multi-source heterogeneous data including device data, user behavior data and third party data, and form data assets by preprocessing the multi-source heterogeneous data. An intelligent decision-making layer is used to identify the current user life cycle stage based on an intelligent model matrix and the data assets, predict the transition stage of the user life cycle, and generate precise marketing strategies and channel selection schemes for each stage. A business application layer is used to convert the precise marketing strategies and channel selection schemes into specific marketing tasks, monitor marketing effect data, and feed back the marketing effect data to the data support layer and the intelligent decision-making layer.

[0011] Based on the same inventive concept, the present application also provides an electronic device comprising a memory and a processor.

[0012] Based on the same inventive concept, the present application also provides a computer readable storage medium having computer executable instructions stored therein, which implement the functions of the above intelligent marketing system when executed.

[0013] Compared with the prior art, the present application has the following advantages: (1) Through the multi-source data collection network and lake warehouse integrated processing mechanism, the island problem of multi-source heterogeneous data is solved, the high-quality integration and efficient utilization of data assets are realized, and the foundation of precision marketing is laid; (2) The intelligent model matrix is used to realize the accurate identification of the user life cycle stage and the scientific prediction of the stage transition, realize the dynamic fine management of the user life cycle, and break through the limitation effect of static grouping; (3) A marketing closed loop of data input, decision output, effect feedback and model optimization is constructed, so that the intelligent marketing system can drive the algorithm model to self-optimize based on the real-time feedback of the marketing effect data, ensure that the marketing strategy always keeps pace with the user dynamics and behavior changes, and improve the operation efficiency and decision scientificity.

[0014] Other features and advantages of the present application will be set forth in the following description, and, in part, will be apparent from the description, or can be learned by practice of the present application. The objects and other advantages of the present application can be achieved and obtained by the structure indicated in the specification, claims and drawings. BRIEF DESCRIPTION OF DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0016] Figure 1 The structure schematic diagram of the intelligent marketing system in the embodiments of the present application is shown; Figure 2 The function schematic diagram of the data development governance operation integrated platform in the embodiments of the present application is shown; Figure 3-1 The business process of the introduction period and the development period of the user life cycle in the embodiments of the present application is shown; Figure 3-2 The business process of the growth period and the loss period of the user life cycle in the embodiments of the present application is shown. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical scheme and advantages of the embodiments of the present application more clear, the technical scheme in the embodiments of the present application will be clearly and completely described in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0018] As Figure 1 shown, the embodiment of the present application provides a smart marketing system based on big data analysis and user life cycle, comprising: a data support layer, an intelligent decision-making layer, and a business application layer.

[0019] The system adopts a three-level architecture system, and each level realizes full-link data interaction through a distributed message bus to form a complete smart marketing closed loop of "data input→algorithm decision→application execution→effect feedback→model iteration". The collaborative logic of the three-layer architecture system is that the data support layer provides high-quality data assets, the intelligent decision-making layer outputs executable marketing decisions based on the data assets, the business application layer converts the marketing decisions into marketing tasks and feeds back the marketing effects, and drives the data support layer and the intelligent decision-making layer to continuously optimize.

[0020] In the embodiment of the present application, the data support layer is used to collect multi-source heterogeneous data, including device data, user behavior data, and third-party data, and form data assets by preprocessing the multi-source heterogeneous data.

[0021] The intelligent decision-making layer is used to identify the current user life cycle stage based on the intelligent model matrix and the data assets, and predict the transition stage of the user life cycle, and generate precise marketing strategies and channel selection schemes for each stage.

[0022] The business application layer is used to convert the precise marketing strategies and channel selection schemes into specific marketing tasks, monitor marketing effect data, and feed back the marketing effect data to the data support layer and the intelligent decision-making layer.

[0023] According to a preferred embodiment, the data support layer comprises a data collection network and a lake-warehouse integrated data processing platform, which provides high-quality data support for the entire marketing system. At the same time, the system supports adding new data sources, sets a dynamic interface adaptation module, and newly added data can be flexibly accessed.

[0024] The data collection network is used to collect the multi-source heterogeneous data, wherein the device data includes device running state data and device basic information, the user behavior data is user behavior data collected according to user end APP and applet behavior burying, and the third-party data is industry-related data obtained by interacting with external systems through API interfaces.

[0025] Taking the battery replacement service as an example, the data collection network is responsible for collecting various data related to the battery replacement user, including: 1) equipment data: high-frequency collection of running state data of the battery replacement equipment, such as voltage, current, battery temperature during the battery replacement process, and also including the geographic position of the battery replacement station, the equipment ID and other basic information; 2) user behavior data: through behavior tracking in the APP and the mini program used by the user, collecting data such as user login records, orders, payment information, package selection, etc.; 3) third-party data: through API interface and external system data interaction, obtaining relevant industry data, etc.

[0026] The lake-warehouse integrated data processing platform is used for preprocessing the multi-source heterogeneous data.

[0027] The lake-warehouse integrated data processing platform includes a data cleaning module, a data fusion module, and a hierarchical storage module.

[0028] The data cleaning module is used for detecting and isolating abnormal data in the multi-source heterogeneous data (such as identifying voltage jump data caused by sensor failure) by using an improved isolation forest algorithm, and removing duplicate data (such as user duplicate registration information) by using a DBSCAN clustering algorithm based on dynamic threshold.

[0029] The data fusion module is used for implementing deep fusion of structured data and unstructured data based on a cross-modal fusion algorithm based on an attention mechanism, such as fusing structured battery replacement order data and unstructured user feedback text data, wherein the text data is semantically encoded by using a BERT model, the structured data is converted into a low-dimensional feature vector by using an embedding layer, and the correlation weight of the two types of data is captured by using a multi-head attention layer.

[0030] The hierarchical storage module is used for dividing the cleaned and fused data into hot data, warm data and cold data according to the access frequency, and storing them in a Redis cluster, an HBase database and an object storage system respectively to form data assets.

[0031] In the lake-warehouse integrated storage architecture, the hot data (user behavior data in the last 7 days, real-time battery replacement orders, etc.) is stored in the Redis cluster to support millisecond-level query; the warm data (battery replacement records in the last 30 days, user package information, etc.) is stored in the HBase database to meet high-concurrency read and write; and the cold data (historical data over 30 days) is archived to the object storage system to realize batch backtracking analysis through SparkSQL.

[0032] Preferably, the data collection network and the lake-warehouse integrated data processing platform are integrated into a data development governance and operation integrated platform. Figure 2 A functional schematic diagram of the data development governance and operation integrated platform is shown, which is configured with: The data acquisition module collects heterogeneous data from multiple sources, including device data, user data, and third-party data, and completes the unified access of these heterogeneous data sources, serving as the starting point for data management.

[0033] The data storage module divides storage schemes according to data type, and stores data through data infrastructure (including CDH system, big data platform, MPP system) and data lake (used to store structured data, semi-structured data, and unstructured data).

[0034] The data planning and development management module is responsible for the entire process of data management from "demand to implementation," and it is divided into four sub-modules: Data planning: Developing data management rules, including standards, architecture standards, and security measures; Requirements management: Complete the entire process of requirements control, including requirements entry, requirements design, requirements development, and requirements deployment; Data integration: Through steps such as real-time acquisition, incremental acquisition, batch synchronization, and data source management, raw data is transformed into usable data; Data development: covering offline development, real-time development, project management, real-time integration and other development stages, realizing the processing of data from "raw data → usable data"; Data Operations and Maintenance: Ensure the stable operation of data flow through 360° monitoring, task monitoring, and alarm management.

[0035] The data governance module, which ensures data quality and value, is divided into five main areas: Master data management: manages master data standards, master data permissions, etc. Data asset management: managing metadata, data assets, data value, etc., to achieve asset-based inventory of data; Data quality management: Ensure data accuracy through quality monitoring, alarm handling, and quality reporting; Data security management: Ensure compliant data use through classification and grading, security gateways, and dynamic data masking.

[0036] In this embodiment, the intelligent decision-making layer takes user lifecycle management as its core and forms an intelligent model matrix integrating stage identification, transition prediction, and marketing strategy based on the feature data provided by the data support layer. This matrix directly drives the marketing decisions of the business application layer and achieves closed-loop optimization based on the feedback from the business application layer.

[0037] According to a preferred embodiment, the intelligent model matrix in the intelligent decision-making layer includes: a user lifecycle model module for managing and predicting user lifecycle stages, a precision marketing algorithm module for generating marketing strategies and channel solutions, and an algorithm coordination scheduling module for coordinating the operation of each model.

[0038] Specifically, the user life cycle model module comprises a stage identification model and a stage transition prediction model.

[0039] The stage identification model is configured to be based on an improved XGBoost algorithm and to fuse a business rule hard constraint to take multi-dimensional and quantifiable indexes strongly associated with each stage of the user life cycle as input and output a label of a current life cycle stage of the user.

[0040] The stage transition prediction model is configured to adopt a hybrid architecture of a time series convolution network (TCN) and a bidirectional long short-term memory network (BiLSTM), combine a business flow rule, and predict a probability of transition of the user to a subsequent life cycle stage, wherein the TCN network is configured to capture short-term behavior mutation features of the user, and the BiLSTM network is configured to learn a long-term behavior evolution trend of the user.

[0041] For example, according to actual operation experience of the battery swap business, the business stages of the user life cycle are determined, each stage has clear business features and determination rules, and the intelligent decision layer constructs the stage identification model and the stage transition prediction model based on the same.

[0042] Specifically, a working process of the stage identification model is as follows: Feature engineering: the business rule is converted into a calculable quantifiable feature, the core is to extract multi-dimensional features strongly associated with each stage of the user life cycle from the whole process of the battery swap business, including user attributes, user behaviors and other features, the abstract business determination rule is converted into a quantifiable index recognizable by the algorithm, and the model input is deeply bound with the business logic.

[0043] Model construction: the improved XGBoost algorithm is adopted, and the deep fusion of the business rule is performed, the XGBoost algorithm is selected as the basic model because it is good at processing structured features and supports a custom objective function, the non-linear association between the features and the stage is learned through the decision tree splitting logic, the regularization term is supported to control the model complexity, and overfitting is avoided.

[0044] Model execution: in order to be closer to the business, the business rule hard constraint is embedded, that is, the stage definition that cannot be broken through in the business flowchart is directly written into the model, and the model is forced to learn the determination rule conforming to the business logic.

[0045] Model optimization: By optimizing the objective function and parameter settings, the accuracy of the model in identifying business stages is improved. The specific solutions are as follows: 1) Add a penalty term to the objective function to impose a high weight penalty on "prediction that violates stage time logic", set a high penalty coefficient to force the model to reduce such errors, and set the penalty term according to the flow order of the business stage to prohibit reverse transition prediction; 2) Set business adaptation parameters, such as tree depth = 7, to adapt to the complexity of 5-stage classification; learning rate = 0.08, to ensure slow iteration and sufficient learning of business rules; L1 regularization = 0.02, to filter weakly correlated features by sparsifying feature weights.

[0046] At the same time, in order to completely eliminate the conflict between model prediction and business process diagram, a rule engine is introduced to calibrate the model output again, ensuring 100% compliance with business logic.

[0047] The stage transition prediction model is based on the analysis of user behavior patterns to predict the likelihood of users transitioning from the current life cycle stage to the subsequent stage, providing a forward-looking basis for marketing intervention, and ensuring wide applicability in the battery swap business scenario.

[0048] Specifically, the workflow of the stage transition prediction model is as follows: Time series feature processing: Based on the time evolution of user behavior, time series features that fit the business cycle are constructed to retain behavior trends and fluctuations: 1) According to the user behavior cycle of the battery swap business, flexible time granularity is selected; 2) Extract general features related to stage transition, including user behavior features, consumption features, and interaction features; 3) Feature vector transformation, i.e. converting user behavior data into time series vectors recognizable by the model, standardizing various features, and concatenating them into sequence vectors in chronological order; 4) Extract local features through sliding windows to enhance the model's ability to capture behavior mutations.

[0049] Model architecture: A hybrid architecture of time convolution network TCN and bidirectional long short-term memory network BiLSTM is adopted, which takes into account the capture of short-term behavior fluctuations and long-term trends, and adapts to the complex time series characteristics of user behavior in the battery swap business.

[0050] Among them, the TCN network is used to capture short-term behavior patterns, extract local time series features of user behavior through convolution operation, adapt to short-term signals before stage transition, identify short-term mutations of user behavior, and only use historical data to use causal convolution and residual connection to avoid future information leakage and improve the model's learning ability of local rules. The BiLSTM network is used to capture the long-term behavior trend of the user, learn the long-term dependence relationship of the user behavior through the bidirectional memory unit, adapt to the stage transition that needs long-term observation, combine the forward and backward time series information, and fully capture the evolution law of behavior to avoid trend misjudgment caused by single time direction.

[0051] Predictive output and business application: the model outputs the transition probability of users to each subsequent stage, providing a general decision basis for marketing intervention, and when the probability of the core transition path reaches a threshold, early warning information is output, reserving a time window for marketing intervention.

[0052] As shown in Figure 3-1 and Figure 3-2 , the intelligent marketing system is configured to perform stage operation logic based on the user life cycle of the introduction period, development period, growth period and loss period, A, B, C and D are connection points.

[0053] The operation logic of the introduction period is configured to respond to new user registration or first order behavior, identify and mark new users entering the introduction period according to preset rules, and automatically transfer new users to the development period after meeting the preset conditions. For new users from different channels, match the corresponding introduction identification rules.

[0054] Among them, the development period includes a new user activation period and an old user recall period.

[0055] The operation logic of the new user activation period is configured to respond to the new user not completing the activation behavior within a preset time, trigger the activation strategy to promote the completion of the first purchase, and successfully transfer the new user to the growth period. Illustratively, after the new user enters the development period, "activation identification" (such as not ordering within 7 days after registration) is triggered, and the first purchase is promoted through "activation strategy 1 (first order coupon)", "activation strategy 2 (new person gift)" and the like. If the activation is successful, it enters the growth period.

[0056] The operation logic of the old user recall period is configured to respond to the old user appearing a loss warning signal, trigger the recall strategy to awaken the old user, and successfully transfer the old user to the growth period. Illustratively, when the old user appears a "loss warning" (such as 30 days of inactivity), the old user is awakened through "recall strategy 1 (exclusive coupon)", "recall strategy 2 (benefit reminder)" and the like. If the recall is successful, it enters the growth period.

[0057] The operation logic of the growth period is configured to respond to the completion of the first purchase by new users and the successful recall of old users, trigger the active strategy to promote the repeat purchase by monitoring the active state of new and old users, and if the user activity decreases, it will be transferred to the loss period. Illustratively, after the new user completes the first purchase or the old user recall is successful, it enters the growth period; through "active identification" (such as active in the last 15 days but not repeat purchase), "active strategy (repeat purchase / points incentive)" is triggered, if the user is continuously active, it will be maintained in the growth period; if the activity decreases, it will enter the loss period.

[0058] The operation logic of the churn period is configured to perform a plurality of recovery strategies with increasing intensity in a gradient manner in response to the user being inactive for a duration exceeding a preset threshold, and if the recovery is successful, the user is returned to the growth period, and if the recovery fails, the user is marked as a churned user. For example, if the user is inactive for more than the threshold (e.g., not logged in for 60 days), the user enters the churn period, and the recovery strategies are performed in a gradient manner, for example, from “recovery strategy 1 (SMS reminder)” to “recovery strategy 5 (dedicated customer service)”, the reach intensity is gradually upgraded, if the recovery is successful, the user returns to the growth period, and if the recovery fails, the user is marked as a churned user.

[0059] Preferably, the stage transition prediction model considers business rule fusion, embeds the stage flow rule of the battery replacement business, and constrains the model output to comply with the actual operation logic, explicitly allows the allowed transition path (such as development period→growth period→churn period), and prohibits unreasonable transitions in the reverse or cross-stage direction (such as growth period→development period). The rule engine is used to calibrate the model output, so as to ensure that the prediction result is consistent with the stage flow logic defined by the business.

[0060] In the embodiments of the present application, the precise marketing algorithm system takes the user life cycle stage as the core to realize the precise placement and maximum effectiveness of the marketing resources of the battery replacement business. The system design deeply adapts to the user interaction scene and operation target of the battery replacement business, and ensures that the strategy and channel are always adapted to the user demand and business rules.

[0061] According to a preferred embodiment, the precise marketing algorithm module comprises a marketing strategy recommendation model and a channel selection optimization model.

[0062] The marketing strategy recommendation model is used to construct a user life cycle stage-specific strategy library, based on a reinforcement learning algorithm, with marketing conversion rate and user life cycle value improvement as reward targets, and dynamically optimizing the decision combination to output marketing strategies matched with each stage of the user life cycle.

[0063] The user life cycle stage-specific strategy library presets marketing strategies adapted to the battery replacement business for each life cycle business stage, and constructs a mapping relationship between the strategy and the stage. The decision dynamic optimization mechanism is based on a reinforcement learning algorithm, with marketing conversion rate and user life cycle value (LTV) improvement as reward targets, and optimizes the strategy combination in real time to adapt to changes in user behavior and adjustments in business targets.

[0064] The channel selection optimization model is used to use a multi-armed bandit (MAB) algorithm to evaluate the marketing effect of each touch channel in real time, and dynamically adjust the channel touch weight to improve the efficiency of marketing resource placement.

[0065] Preferably, the response rate difference of users in different life cycle stages to the touch channel is analyzed, a channel preference model is constructed, and a multi-arm bandit (MAB) algorithm is used to monitor the marketing effect of each channel in real time, dynamically adjust the channel touch weight, preferentially select high-response and high-conversion channels, and improve the efficiency of marketing resource investment.

[0066] In the embodiments of the present application, the algorithm coordination mechanism is the core support for guaranteeing the efficient operation of the whole link of user life cycle stage identification, stage transition prediction and strategy recommendation, ensuring that the intelligent decision-making layer modules respond to the battery swap business demand in coordination, and realizing seamless connection from data input to strategy output. The algorithm coordination mechanism is oriented to business processes, builds an orderly collaboration and automatic scheduling mechanism among algorithm modules, ensures the coherence of data flow and task execution, and adapts to the real-time and periodic needs of the battery swap business.

[0067] According to one preferred embodiment, the algorithm coordination scheduling module comprises a process linkage scheduling unit, a self-optimization unit and a business rule adaptation module.

[0068] The process linkage scheduling unit is configured to support two modes of fixed-period start whole-link task timing scheduling and event-driven immediate start associated scheduling, and push the algorithm results, including the labels of user life cycle stages, stage transition probabilities and marketing strategy instructions, to the business application layer in real time through standardized interfaces.

[0069] The process linkage logic of the process linkage scheduling unit is to build a modular linkage framework based on the core link of the business scenario, orderly connect each algorithm module according to the dependency relationship, provide high-quality input for feature generation from data cleaning, support the inference of the stage identification model from feature generation, drive the output of the strategy recommendation model from the stage identification result and the transition prediction probability, and form close linkage of upstream output and downstream input.

[0070] The scheduling mechanism of the process linkage scheduling unit is to adapt to the trigger needs of different business scenarios by using flexible scheduling strategies, start the whole-link task timing scheduling in a fixed period for periodic business, ensure the continuous tracking of user behavior patterns, or start the associated task immediately for scenarios with high real-time needs, and realize real-time response of the strategy.

[0071] The instruction delivery mechanism of the process linkage scheduling unit is that the algorithm results such as stage labels, transition probabilities and marketing strategy instructions output by the intelligent decision-making layer are pushed to the business application layer in real time through standardized interfaces, so that the business application layer can execute marketing actions based on the latest algorithm decisions, and avoid strategy failure caused by information lag.

[0072] The self-optimization unit is configured to dynamically adjust parameters of each model in the intelligent model matrix based on real-time user behavior and feedback data of the business application layer in an online learning mode, to adapt to short-term user behavior fluctuations, and automatically start full data retraining when the model performance index continuously falls below the business expectation.

[0073] The self-optimization unit ensures that the algorithm model always adapts to changes in user behavior and adjustments in operating rules of the battery swap business, and maintains long-term effectiveness through a dynamic iteration and rule synchronization mechanism. Further, based on marketing effect data fed back by the business application layer, a continuously iterative optimization link is constructed, for example, an online learning mode is adopted to dynamically adjust model parameters in real time by absorbing newly generated user behavior and strategy feedback data, so that the model can quickly adapt to short-term behavior fluctuations. Further, when the model performance index continuously falls below the business expectation, full data retraining is automatically started to optimize the model structure and parameters in combination with the evolution law of the business scenario, thereby ensuring long-term prediction accuracy. Further, the weight of each model is dynamically adjusted in the direction of the core target of the business.

[0074] The business rule adaptation module is configured to automatically convert new rules into constraint conditions recognizable by algorithm models in the intelligent decision-making layer based on adjustment instructions of business operation rules, and synchronize rule change events to all-link related modules of the data support layer and the intelligent decision-making layer, and start automated verification after rule synchronization to ensure that the adapted results are consistent with the business expectation by comparing the differences in model output under new and old rules.

[0075] The business rule adaptation module realizes real-time synchronization of the intelligent decision-making layer and business operation rules through a rule engine, ensuring that technical decisions do not deviate from business reality. When the core rules of the business are adjusted, the rule engine automatically converts new rules into constraint conditions recognizable by algorithms. Further, the rule change is synchronized to all-link modules, the feature engineering updates the calculation logic of business-related features, the stage recognition model adjusts the judgment boundary, and the strategy recommendation model updates the strategy library and optimization target, to ensure that the whole process from data input to strategy output conforms to the new rules. Further, after rule synchronization, automated verification is started to ensure that the adapted results are consistent with the business expectation by comparing the differences in model output under new and old rules, thereby avoiding deviations in the rule transmission process.

[0076] In the embodiments of the present application, the business application layer converts the marketing strategies output by the intelligent decision-making layer into specific marketing tasks, monitors the execution effect of the marketing tasks, and feeds back the marketing effect data to the data support layer and the intelligent decision-making layer.

[0077] According to a preferred embodiment, the business application layer includes a user insight center, an intelligent marketing execution engine, and an effect evaluation and iteration center.

[0078] The user insight center is used for integrating and visualizing display of labels including user life cycle stages, stage transition probabilities, and user base attributes, behavior preference data, constructing a user panoramic portrait (for example, a 35-year-old rider, an active user, and a preference for night battery replacement), and outputting a target user package.

[0079] The intelligent marketing execution engine is used for automatically matching a target user group, formulating a differentiated marketing task, and scheduling a touch channel according to the target user package based on instructions of a marketing strategy output by the intelligent decision layer. The intelligent marketing execution engine automatically generates a marketing task according to the strategy instructions output by the intelligent decision layer, that is, matching a target user package, executing strategy content, scheduling a touch channel, and the like. Real-time optimization logic exists in strategy execution, and differentiated execution rules are carried out for different user groups.

[0080] The effect evaluation and iteration center is used for monitoring marketing activity data, including click rates, conversion rates, and value changes of user life cycles, calculating contribution degrees of each marketing touch point through a multi-touch weight distribution attribution model, generating an evaluation report, and feeding back the evaluation report to the data support layer and the intelligent decision layer in real time.

[0081] The present application embodiment realizes a closed loop through high-quality processing of the data support layer, dynamic decision of the intelligent decision layer, and effect feedback of the business application layer, overcomes defects such as weak data integration capability, static user management, data link breakage, and poor scene adaptability in the prior art, and realizes precision, automation, and sustainable optimization of intelligent marketing. The beneficial effects of the present application embodiment are as follows: (1) Solving multi-source data processing quality and efficiency problems, and improving data asset value In the prior art, traditional marketing relies on single static data. Although some schemes involve multi-platform data, the processing is rough. The present application embodiment realizes multi-source data collection network and lake-warehouse integrated processing mechanism, improves the accuracy of data by accurately cleaning abnormal data and repeated data based on an improved isolated forest algorithm and a dynamic threshold DBSCAN clustering, adopts a cross-modal fusion algorithm with an attention mechanism to fuse structured order data and unstructured user feedback text, solves the problem of multi-source heterogeneous data island, and supports millisecond-level query and batch backtracking analysis through hierarchical storage of hot data, warm data, and cold data, to meet the needs of high concurrency and historical data reuse.

[0082] (2) Realizing dynamic management of user life cycle and precise marketing decision, and breaking through the limitation of static grouping The prior art is limited to single-stage identification or lacks stage transition prediction capability, and marketing strategies are disconnected from user dynamic behavior. The intelligent decision layer of the embodiment of the application outputs the current life cycle stage of the user accurately through an intelligent model matrix, and the identification accuracy conforms to business logic. The stage transition prediction model combining the TCN network and the BiLSTM network takes into account short-term behavior fluctuations and long-term trends, and outputs the transition probability in advance to reserve a time window for marketing intervention. The strategy recommendation model and the channel selection model are linked, dynamically generate strategies for different stages, and preferentially select high-response channels to solve the problems of strategy staticization and blind channel selection.

[0083] (3) Build a marketing full-link closed loop to improve execution efficiency and iteration capability Most of the prior art lacks a marketing strategy and effect evaluation closed loop. In the embodiment of the application, the business application layer automatically generates marketing tasks according to the intelligent marketing execution engine, replaces manual coordination, and greatly shortens the marketing response time. The effect evaluation module quantifies the actual impact of the strategy through the attribution model and multi-dimensional index analysis, solves the problems of short-term evaluation and difficult attribution in traditional evaluation, and forms a continuous improvement link of data input-decision output-effect feedback-model optimization to ensure that the system adapts to changes in user behavior and adjustments in business rules.

[0084] (4) Enhance cross-scene adaptability and break through industry limitations Most of the prior art is limited to specific fields and is difficult to migrate to entity service scenarios. In the embodiment of the application, through modular architecture design, only the life cycle stage definition and strategy library need to be adjusted, and more than 80% of the functions can be reused, quickly adapting to multiple scenes such as shared bicycles and charging services, and solving the problems of scene binding and low reusability of traditional solutions.

[0085] Based on the same inventive concept, the application further provides an electronic device, comprising a memory and a processor. The processor is configured to read and execute a computer program stored in the memory to realize the functions of the intelligent marketing system.

[0086] Based on the same inventive concept, the application further provides a computer readable storage medium, wherein the computer readable storage medium stores computer executable instructions. The computer executable instructions realize the functions of the intelligent marketing system when executed.

[0087] Although the application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can modify the technical solutions described in the foregoing embodiments or make equivalent replacements to some technical features, and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the application.

Claims

1. A smart marketing system, characterized in that, The system includes: The data support layer is used to collect multi-source heterogeneous data, including device data, user behavior data and third-party data, and to form data assets by preprocessing the multi-source heterogeneous data. The intelligent decision-making layer is used to identify the current user lifecycle stage based on the intelligent model matrix and the data assets, predict the transition stages of the user lifecycle, and generate precise marketing strategies and channel selection schemes for each stage. The business application layer is used to transform the precision marketing strategy and channel selection scheme into specific marketing tasks, monitor marketing performance data, and feed the marketing performance data back to the data support layer and the intelligent decision-making layer.

2. The system according to claim 1, characterized in that, The data support layer includes: A data acquisition network is used to collect the multi-source heterogeneous data, wherein the device data includes device operating status data and basic device information, the user behavior data is user behavior data collected based on user terminal APP and mini program behavior tracking points, and the third-party data is industry-related data obtained through API interface interaction with external systems; The integrated lake-warehouse data processing platform is used for preprocessing the multi-source heterogeneous data; The integrated lake warehouse data processing platform includes: The data cleaning module is used to detect and isolate abnormal data in the multi-source heterogeneous data using an improved isolated forest algorithm, and to remove duplicate data using a DBSCAN clustering algorithm based on dynamic thresholds. The data fusion module is used for attention-based cross-modal fusion algorithms to achieve deep fusion of structured and unstructured data; The tiered storage module is used to divide the cleaned and merged data into hot data, warm data, and cold data according to the access frequency, and store them in Redis cluster, HBase database, and object storage system respectively to form data assets.

3. The system according to claim 1, characterized in that, The intelligent model matrix in the intelligent decision-making layer includes: a user lifecycle model module for managing and predicting user lifecycle stages, a precision marketing algorithm module for generating marketing strategies and channel solutions, and an algorithm coordination scheduling module for coordinating the operation of each model.

4. The system according to claim 3, characterized in that, The user lifecycle model module includes: The stage identification model is based on the improved XGBoost algorithm, which integrates hard constraints of business rules, and takes multi-dimensional and quantifiable indicators that are strongly correlated with each stage of the user lifecycle as input, and outputs the label of the current stage of the user lifecycle. The stage transition prediction model uses a hybrid architecture of temporal convolutional network (TCN) and bidirectional long short-term memory network (BiLSTM) to predict the probability of a user transitioning to a subsequent lifecycle stage, combined with business flow rules. The TCN network is used to capture the short-term behavioral mutation characteristics of users, while the BiLSTM network is used to learn the long-term behavioral evolution trends of users.

5. The system according to claim 4, characterized in that, The system is configured to execute phased operational logic based on the user lifecycle, including the introduction, development, growth, and churn phases. The operational logic of the introductory phase is configured to respond to new users completing registration and first order behavior, identify and mark new users to enter the introductory phase according to preset rules, and automatically transfer new users to the development phase after the preset conditions are met; The development period includes a new user activation period and an old user recall period; The operational logic for the new user activation period is configured to trigger an activation strategy to encourage a new user to complete their first purchase if they fail to activate within a preset time. Once successful, the new user will be transferred to the growth period. The operational logic for the old user recall period is configured to respond to the old user churn warning signal, trigger the recall strategy to wake up the old user, and after success, transfer the old user to the growth period. The operational logic during the growth phase is configured to respond to new users completing their first purchase and old users being successfully recalled by monitoring the activity status of new and old users to trigger an activity strategy to promote repeat purchases. If user activity declines, the user will be transferred to the churn phase. The operational logic for the churn period is configured to respond to a user's inactivity time exceeding a preset threshold by executing multiple recovery strategies with increasing intensity in a tiered manner. If the recovery is successful, the user is returned to the growth phase; if it fails, the user is marked as a churned user.

6. The system according to claim 4, characterized in that, The precision marketing algorithm module includes: The marketing strategy recommendation model is used to build a dedicated strategy library for each stage of the user lifecycle. Based on reinforcement learning algorithms, it dynamically optimizes decision combinations with marketing conversion rate and user lifecycle value improvement as reward objectives, so as to output marketing strategies that match each stage of the user lifecycle. The channel selection optimization model uses the Multi-Armed Slots (MAB) algorithm to evaluate the marketing effectiveness of each reach channel in real time and dynamically adjust the channel reach weight to improve the efficiency of marketing resource allocation.

7. The system according to claim 6, characterized in that, The algorithm-coordinated scheduling module includes: The process linkage scheduling unit is configured to support two modes: timed scheduling of full-link tasks that start at a fixed period and instant-start associated scheduling through event-driven methods. It pushes algorithm results, including user lifecycle stage tags, stage transition probabilities, and marketing strategy instructions, to the business application layer in real time through a standardized interface. The self-optimizing unit is configured to dynamically adjust the parameters of each model in the intelligent model matrix based on real-time user behavior and feedback data from the business application layer, using an online learning mode to adapt to short-term fluctuations in user behavior, and automatically start full data retraining when the model performance indicators are consistently lower than business expectations. The business rule adaptation module is configured to automatically convert new rules into constraints that can be recognized by the algorithm model in the intelligent decision layer based on adjustment instructions of business operation rules. It also synchronizes rule change events to the full-link related modules of the data support layer and the intelligent decision layer, and starts automatic verification after rule synchronization. By comparing the differences in model output under the new and old rules, it ensures that the adapted results are consistent with business expectations.

8. The system according to claim 7, characterized in that, The business application layer includes: The User Insight Center is used to integrate and visualize user lifecycle stage tags, stage transition probabilities, user basic attributes, and behavioral preference data to build a comprehensive user profile and output target user packages. The intelligent marketing execution engine is used to automatically match target user groups, formulate differentiated marketing tasks, and schedule outreach channels based on the instructions of the marketing strategy output by the intelligent decision-making layer and the target user package. The Performance Evaluation and Iteration Center is used to monitor marketing campaign data, including click-through rate, conversion rate, and changes in user lifetime value. It calculates the contribution of each marketing touchpoint through a multi-touchpoint weight allocation attribution model, generates an evaluation report, and feeds the evaluation report back to the data support layer and the intelligent decision-making layer in real time.

9. An electronic device, characterized in that, include: Memory, processor; The processor is configured to read and execute the computer program stored in the memory to implement the functions of the system according to any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed, perform the functions of the system according to any one of claims 1-7.

Images